​

LASSO Lab, https://lassolab.org/

Director: Toby Dylan HOCKING

Learning Algorithms, Statistical Software, Optimization

Machine learning adapts to patterns in different data

• Two image classification data sets: digits and clothing
• Wide range of other applications: cybersecurity, robotics, �climate/carbon modeling, genomics, neuroscience, medicine, microbiome.
• ​

Cross-validation proves accurate predictions in new data

• Software: https://github.com/tdhock/mlr3resampling
• Example: predicting autism diagnosis.
• Can we train �on 2019 �and predict �accurately �on 2020?�(Compare �Other and�Same).

Learning algorithms have different accuracy and speed

• Image segmentation/classification, and audio classification data sets.
• xgboost is slowest to train on all four data sets, most accurate in aztrees3, EMNIST.
• Other algorithms are faster and more accurate in vowel, waveform.
• Need to try a variety of algorithms for optimal prediction in any data.

​

Interpretable machine learning: feature selection

• Example: predicting autism diagnosis, which features contribute?
• 364 features to consider; very accurate using only 21 with linear model.
• Tibshirani 1996 LASSO paper

Interactive data visualization

• Examples: https://animint.github.io/gallery/
• Click on data points to�see more details
• Useful for exploring and�presenting complex data�and models
• Similar to Tableau and�PowerBI but with R API�and more advanced �(multiple selection,�multiple layers, etc).

Conclusion: LASSO Lab strengths

• Learning Algorithms: neural networks, boosting, interpretable linear models, etc, training in parallel, evaluating prediction accuracy using cross-validation.
• Statistical Software: free/open-source libraries in R/Python/C/C++/JavaScript for machine learning and data visualization.
• Wide range of applications: cybersecurity, robotics, satellite/sonar imagery, �climate/carbon modeling, genomics, neuroscience, medicine, microbiome.
• Contact: toby.dylan.hocking@usherbrooke.ca